| The digital representation of three-dimensional objects is generally a little cloud,voxel,grid,etc.Point cloud is the most widely used digital representation of threedimensional objects because it retains the original collected data and is easy to store and use.In reality,due to problems such as occlusion,lighting,equipment,etc.,it is easy to cause large defects and incompleteness in the collected three-dimensional point cloud data.With the development of three-dimensional computer vision,the completion of missing point clouds has become the basic core work of virtual reality,digital twin,and smart city.Generally speaking,the point cloud completion task usually starts with the incomplete,local and incomplete point cloud,and then obtains the complete point cloud.The intermediate operation is usually to use various algorithms and shape conversion.The point cloud completion task is of great significance to the development of autodriving and virtual reality,which are often accompanied by a large number of incomplete and local point clouds.A good point cloud model can make these two work better.Current research on point cloud completion is based on in-depth learning.Most of the research is based on point method,while a few work is based on voxel method.This paper presents a point cloud completion model that fuses multimodal information because it incorporates image information,residual edge convolution,channel attention mechanism,and spatial attention mechanism.Local detail recovery is good for three-dimensional point clouds.Specifically,this paper is divided into two parts:(1)Designed a point cloud completion model that combines a dual-attention mechanism with a dynamic convolution neural network.The model first constructs edge features from residual edge convolution blocks to point clouds,then brings center and domain points into the computed enhanced edge features to enhance the associated information between center and domain points.Secondly,the spatial attention mechanism is inserted into the model as a module to enhance the relationship between domain points,extract and preserve the characteristics between them.Then,channel attention mechanism is used when aggregating local features,i.e.combining maximum pooling with average pooling to assign different channels different weights to achieve the goal of emphasizing useful channels and suppressing useless ones,emphasizing that useful channels inhibit useless channels by giving different weights to different channels.The decoder decodes the aggregated point cloud features into complete point clouds using the folding operation constrained by sparse point clouds.The results of comparison and ablation experiments on Shape Net dataset-PCN dataset,an open dataset provided by Stanford University,show that the point cloud remediation method proposed here,which combines two-channel attention mechanism and residual edge convolution,improves the CD value and machine learning index in most categories of three-dimensional objects compared with the existing mainstream point cloud remediation methods.This experiment fully demonstrates the validity of this method.(2)For the sparsity of point clouds,a multimodal point cloud completion model based on single view image information is proposed.First,the encoder in(1)is used to extract the features of point clouds,then single-view images are mapped to point cloud modes through a series of convolution and deconvolution blocks of different scales.The point cloud is reconstructed using the reconstructed point clouds and residual defect clouds as constraints for point cloud generation.Finally,the aggregated point cloud features are re-decoded in the constraints to get the full point cloud using the collapse operation.The ablation and comparison experiments on the Shape Net dataset demonstrate that the point cloud completion model based on point cloud and single view modal information fusion proposed by this method is more uniform in structure and more advanced in evaluation index than the previous single mode point cloud completion model based on learning.At the same time,the experimental results also show that using image information as a supplement to point cloud information can better complete the point cloud completion task. |
